vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

added support for 70b llama 2

Browse source
This commit is contained in:
Concedo 2023-07-24 15:20:18 +08:00
parent 993ba3b026
commit 94499dba25
1 changed files with 5 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

5
llama.cpp
View file

@ -1045,6 +1045,11 @@ static void llama_model_load_internal(
// LLaMAv2 // LLaMAv2
// TODO: temporary until GGUF // TODO: temporary until GGUF
//patch for llama2 gqa
if (model.type == e_model::MODEL_65B && hparams.n_mult >= 4096) {
fprintf(stderr, "%s: Applying KCPP Patch for 70B model, setting GQA to 8\n", __func__);
n_gqa = 8;
}
LLAMA_ASSERT(hparams.n_head % n_gqa == 0); LLAMA_ASSERT(hparams.n_head % n_gqa == 0);
hparams.n_head_kv = hparams.n_head / n_gqa; hparams.n_head_kv = hparams.n_head / n_gqa;
if (model.type == e_model::MODEL_65B && n_gqa == 8) { if (model.type == e_model::MODEL_65B && n_gqa == 8) {